AIF-C01 Question 115
Select 2A company wants to customize a foundation model for their specific use case of analyzing customer reviews and extracting sentiments. They are evaluating pre-training, fine-tuning, in-context learning, and retrieval-augmented generation (RAG) approaches. Which approaches are likely to result in the lowest cost while still achieving reasonable performance for their task?
- A
Pre-training a new foundation model from scratch
- B
Fine-tuning an existing foundation model on labeled customer reviews
- C
Using in-context learning by providing customer reviews as examples in the input prompt
- D
Implementing retrieval-augmented generation (RAG) to provide customer review data as context during inference
Show answer and explanation
Correct answers: C, D
Explanation
For a task like extracting sentiments from customer reviews, in-context learning and retrieval-augmented generation (RAG) are the most cost-effective approaches. Both methods avoid the computational and financial overhead of training or fine-tuning a model. In-context learning utilizes examples directly in the input, while RAG dynamically retrieves relevant information to enhance inference. Pre-training is only suitable for organizations with extensive resources, and fine-tuning, while effective, still incurs higher costs compared to in-context learning or RAG.
- A. Incorrect.
Pre-training a new foundation model from scratch is the most expensive option because it requires significant computational resources, large amounts of data, and time. This is not cost-effective for most business use cases.
- B. Incorrect.
Fine-tuning an existing foundation model can be more cost-effective than pre-training, but it still requires labeled data and computational resources to retrain the model parameters, which can make it relatively expensive.
- C. Correct.
Using in-context learning is a cost-efficient approach since it does not require retraining the model. Instead, examples are provided during inference, leveraging the model's existing capabilities with minimal overhead.
- D. Correct.
Retrieval-augmented generation (RAG) avoids the need for model retraining by dynamically retrieving relevant information during inference. This is cost-effective, especially when dealing with large datasets like customer reviews, as it combines retrieval systems with existing models.